Provisional findings from six NASA recordings

Two recordings consistently favored the globe models over the specified flat disc; four remained unresolved. In the two resolved recordings, the rotating globe also fit better than the still globe under every tested assumption.

The available pilot comparisons point toward globe, rather than equal support for globe and flat. Two further recordings have globe-favoring partial comparisons, and the remaining two have no completed shape comparison. None has a resolved preference for flat. The completed selected-stretch study subsequently found 12 complete globe comparisons, zero complete flat comparisons and 41 unfinished comparisons. All 12 globe results also favor rotation.

The pilot used the same saved IMU and receiver observations for all three models. Each recording had four matched cases: constant gyro-offset allowances of ±0.1 or ±1 degree/hour, with Earth-rate removal either absent or fitted. There were 72 model fits in total.

Model comparisons and unfinished calculations

40 fits converged; 32 did not. No fit returned a numerical or framing error. There were 73 charged starts, including one interrupted attempt; every completed fit was retained. A fit adjusts the permitted instrument and motion parameters to match the recording. Convergence means those adjustments passed the numerical stopping checks. It does not mean the model describes the recording well. A converged flat-disc fit can still have much larger errors than a converged globe fit. The model comparison uses the quality of the matched fits, rather than counting how many calculations finished. Convergence also does not establish a global optimum or validate the error assumptions.

The audited results identify the exact original filenames. No recording had a resolved preference for the flat disc. The four unknown results are not votes for either shape: at least one required optimization remained unfinished. Rotation is withheld for them rather than interpreted from incomplete comparisons.

What do the partial results lean toward?

Thus the available shape pairs in two indeterminate recordings lean toward globe. The other two cannot supply a defensible lean. These partial comparisons retain useful diagnostic information, but do not override the requirement for all necessary fits or turn an unfinished optimization into evidence against a model. Rotation remains withheld.

Numerical completion under different assumptions

The counts below describe solver completion, not support for an Earth model.

Each model had 24 attempted calculations. Twelve rotating-globe calculations, 14 still-globe calculations and 14 flat-disc calculations passed the stopping checks. Those similar counts do not mean the models fit equally well: both recordings with complete comparisons favored globe, and then rotating globe, in every case. They retained the same shape and rotation preferences when the assumed gyro-offset allowance changed or possible Earth-rate removal was introduced. Neither smaller bounds nor additional processing freedom solved the other recordings' convergence problems.

Converged fits used a median of 124 evaluations; unresolved fits used a median of 200. Twenty-three of the 32 unresolved fits used the full 200-evaluation allowance. These differences describe solver behavior, not instrument quality or evidence against an unfinished model. They do not justify dropping the unresolved recordings.

What limits the finding?

Every fit in the two fully converged recordings touched at least one assumed parameter limit. Their preferences therefore need further checks of those limits and model adequacy. The rotating-globe fit's position residuals also differed substantially between recordings: roughly 0.2–4.9 metres horizontally on 26 October 2010, versus 6.7–19.8 metres on 24 September 2015, across the tested cases and horizontal axes. A relative winner can still fit the observations inadequately.

The IMUs are professional survey instruments with published performance information. The outstanding checks concern the remaining calibration offset, exact onboard corrections, integration clock and receiver errors. A raw increment stream can include ordinary sensor filtering without losing slow turning; passing the fit does not prove an absence of onboard corrections. Fit differences have not been converted into validated significance levels. These are six recordings with repeated fits of the same observations, not 72 independent trials.

The larger analysis uses permanently selected high-speed stretches and two wider-offset processing cases. Its findings compare fully completed and unfinished stretches, including duration, speed and IMU record type. Both studies' complete comparisons favor globe and rotation. Their unfinished calculations do not supply equal support for flat Earth.

Full numerical evidence is preserved in fit diagnostics, shape comparisons and rotation comparisons.